Design and optimization of LC oscillators Maria del Mar Hershenson1 Ali Hajimiri California Institue of Stanford University 1 currently with Barcelona Design Technology Sunderarajan S. Mohan1 , Stephen P. Boyd, Thomas H. Lee Stanford University Abstract Vdd We present a method for optimizing and automating component and transistor sizing for CMOS LC oscillators. We observe that the performance measures can be formulated as posynomial functions of the design variables. As a result, the LC oscillator design problems can be posed as a geometric program, a special type of optimization problem for which very efficient global optimization methods have recently been developed. The synthesis method is therefore fast, and determines the globally optimal design; in particular the final solution is completely independent of the starting point (which can even be infeasible), and infeasible specifications are unambiguously detected. We can rapidly compute globally optimal trade-off curves between competing objectives such as phase noise and power. 1 MP1 MP2 L Cvar Cload L Vcontrol Cvar MN1 MN2 Cload Introduction LC oscillators are commonly used in CMOS radio-frequency integrated circuits (RF-ICs) because of their good phase noise characteristics and their ease of implementation [1, 2, 3, 4, 5]. However, no systematic design methodology exists to account for the many specifications (phase noise, power dissipation, voltage swing, tuning range . . . ). Typical design strategies have tried to achieve a low phase noise by using the largest possible inductance value [3], or by using inductors with the lowest possible series resistance [4]. However, an optimal design requires the simultaneous consideration of both active and passive devices. In this paper, we propose a new method for the design of LC oscillators. We illustrate our technique for the specific architecture of Figure 1. The method is easily extensible to other LC oscillator architectures. We formulate the design problem as a special type of optimization problem, called geometric programming (GP). We provide an overview GP in §2 and then outline the design variables in §3. In §4, we obtain the equivalent circuit model for the LC oscillator using the inductor model presented in [6] and simple transistor models. In §5, we use these models and the phase noise model of [5] to show how the design specifications can be posed in a way suitable for geometric programming. This geometric program formulation allows us to obtain globally optimal designs very efficiently, thereby permitting the designer to spend more time exploring design trade-offs rather than sizing inductors and devices. We illustrate our method with design examples in §6 and optimal performance tradeoff curves in §7. Ibias Figure 1: Complementary LC oscillator. 2 Geometric Programming Let f be a real-valued function of n real, positive variables x1 , . . . , xn . It is called a posynomial function if it has the form f (x1 , . . . , xn ) = t X 1k α2k nk ck xα x2 · · · xα , n 1 k=1 where cj ≥ 0 and αij ∈ R. When t = 1, f is called a monomial function. A geometric program (see [7]) has the form minimize f0 (x) subject to fi (x) ≤ 1, i = 1, 2, . . . , m, gi (x) = 1, i = 1, 2, . . . , p, xi > 0, i = 1, 2, . . . , n, (1) where fi are posynomial functions and gi are monomial functions. The most important feature of geometric programs is that they can be globally solved with great efficiency, with no initial point needed, using newly developed interior-point methods [8, 9, 10]. To carry out the designs described in this paper we implemented in MATLAB a very simple (primal barrier) method for solving the convex form of a GP [10]. Despite the simplicity of the method, and the inefficiency of our implementation, all the design problems in this paper were solved under one second on a personal computer. 3 Design variables To design the oscillator we specify the following twelve variables: • Spiral inductor: number of turns n, turn width w, turn spacing s and outer diameter dout . 0-7803-5832-X /99/$10.00 ©1999 IEEE. • Transistors: width (Wn and Wp ) and length (Ln and Lp ). • Varactor: maximum value Cv,max and minimum value Cv,min . g • Transconductance m : A simple model for the transconductance of short-channel devices is [16], gm = µCox W Esat /2, • Load capacitance: Cload . • Bias current: Ibias . Since the thermal noise of the tail current source in the vicinity of the frequency of oscillation does not affect the phase noise of the oscillator due to its differential operation [5], we use Ibias as a design parameter rather than the bias transistor dimensions. 4 4.1 Model for the LC oscillator Inductor model In this paper, we consider planar spiral inductors. One can easily extend the method to other inductors such as bond wires [1], autotransformers [3] and non-standard spiral inductors [2]. A spiral inductor is characterized by the number of turns n, the turn width w, the turn spacing s, and the outer diameter dout . The inductor can be implemented with or without a patterned ground shield (PGS) [11] (a grounded polysilicon shield broken regularly in the direction perpendicular to the inductor current flow). Figure 2 shows a commonly used circuit model for the inductor (with or without PGS) (see [6]). This model is accurate for lightly doped substrates as long as the assumption of a lumped model is valid. In [12], it is shown that the inductor circuit element values are posynomial functions of the design variables (n, dout , w and s). The inductance L can be modeled by a monomial function of the design variables with typical errors of only 3% (see [13]). where µ is the mobility of the carriers in the channel, Cox is the oxide capacitance, W is the transistor width, and Esat is the field at which the carrier velocity reaches half its saturation velocity. g • Output conductance d : We use a simple monomial model for the short-channel transistor output conductance that was used in [15], (3) gd = λI 0.6 L−1 W 0.4 , where λ is a fitting parameter, I is the transistor drain current and L is the transistor channel length. • Capacitances: We model the gate-to-drain (Cgd ), the drainto-bulk (Cdb ) and the gate-to-source (Cgs ) capacitances with posynomial expressions in the transistor width and length (see [16, 15]). 4.4 Tank model −gm,p Cdb,p +Cgs,p Cs L Rs 2 4Cgd,p CL CL Rp Rp Rs Rs Cv Cload −gm,p 4Cgd,p L 1 Rv Cp Rp Figure 2: Simplified inductor model. Varactor model There are several varactor options for frequency tuning [14] such as junction diodes, MOS capacitors and accumulation mode capacitors. In this paper, we use a generic varactor model although models specific to a particular implementation can also be used. The varactor capacitance ranges from some minimum value Cv,min to some maximum value Cv,max . The ratio Cv,max /Cv,min is limited due to physical limitations of the varactor. The varactor quality factor Qv depends on the frequency of operation. For simplicity, we assume a constant quality factor value equal to its minimum value across the tuning range. We model the varactor as an ideal capacitor in series with a resistor Rv = Qv / (Cv ω). We will see that this simplification is not critical since the varactor contributes little to the total oscillator phase noise (see §6). 4.3 Cv Cload 4Cgd,n Cp −gm,n Cdb,n +Cgs,n 4.2 Cdb,p +Cgs,p L Rv 4Cgd,n Rp (2) Transistor model We now describe an analytical model for short-channel devices. More accurate models for transistors that are still compatible for geometric programming are available [15]. −gm,n Cdb,n +Cgs,n Figure 3: Complementary LC oscillator. We model the tank with the equivalent small signal differential mode circuit shown in Figure 3, where the dashed lined is an effective AC ground for differential operation. We now define: • Tank inductance ( L tank )is given by the monomial, Ltank = 2L. • Tank capacitance ( Ctank = C tank )is (4) given by, 1 (4Cgd,n + Cgs,n + Cdb,n + 4Cgd,p 2 +Cgs,p + Cdb,p + CL + Cv ), (5) which is a posynomial function of the design variables since it is a sum of posynomial functions. • Tank load conductance ( g tank ) is given by, gtank = (gd,n + gd,p + gv + gL ) /2, (6) where gv , the effective parallel varactor conductance, and gL , the effective parallel inductor conductance, are given by the posynomial expressions (see [12]), gv = Cvar ω , Qv gL = 1 1 + . Rp (Lω)2 /Rs Since gd,n and gd,p are also given by posynomial functions of the design variables, expression (6) for gtank is a posynomial function of the design variables. Transistor channel thermal noise. In [5], it is shown that the total differential thermal current noise density due to the active devices is given by the posynomial expression, g • Tank effective negative conductance ( neg;tank ) is determined by the transconductance of the cross-coupled transistor pairs. To improve the 1/f 3 corner of the phase noise it is convenient to have a symmetric tank (gm,n = gm,p ) [5]. For symmetric tanks, gneg,tank is given by the monomial expression, gneg,tank = −(gm,n + gm,p )/2 = −gm,n . 5 i2M,dT 1 = ∆f 2 We now show that the design specifications for the LC oscillator can be expressed as either monomial equality constraints, or posynomial inequality constraints, and therefore can be handled by GP. Quiescent power i2M,d ∆f (8) i2M,gT 1 = ∆f 2 The amplitude of the differential voltage across the tank is determined by whether the oscillator is operating in current limited mode or in voltage limited mode. It can be expressed as, Vamp = min{Ibias /gtank , Vdd }. Vamp,min ≤ Ibias Rtank . i2 M,dp ∆f i2 M,gp ∆f MP1 Rp In the 1/f 2 region of the phase noise spectrum, the single sideband phase noise at an offset frequency (foff ) is given by [5] L(foff ) = L P · i2n /∆f 2 qmax , i2M,gn i2M,gp + ∆f ∆f ! , (15) (16) Vdd Phase noise model Γ2rms 2 8π 2 foff (14) where the parameter δ takes a value close to twice γ (see [16]) and Cgs is the gate to source capacitance. Since Cgs is given by a posynomial expression, equation (16) is also posynomial. (9) (10) 4kT γIbias , Esat L = 2 i2M,g 4kT δω 2 Cgs Esat L = , ∆f 5Ibias We can impose a minimum swing (Vamp ≥ Vamp,min ) with the two monomial constraints, Vamp,min ≤ Vdd (13) where for short-channel devices, Since equation (8) is monomial, we can bound the maximum quiescent power with a monomial constraint (P ≤ Pmax ). Differential voltage amplitude , where k is Boltzman constant, T is the temperature in Kelvin and γ ≈ 2 for short channel transistors. This simple model agrees well with measurements [5]. Transistor gate noise. Since the gate noise sources are in parallel with the drain noise sources, they add in the same way [5]. Therefore, the total differential transistor gate current noise density is given by the posynomial expression, The quiescent power is given by the product of the supply voltage Vdd and the bias current Ibias , P = Ibias Vdd . ! where for short-channel devices, (7) Design specifications i2M,dn i2M,dp + ∆f ∆f Rs i2 M,gp ∆f v2 Rp ∆f v2 Rp ∆f Rp v2 Rs ∆f v2 Rs ∆f Rs MP2 i2 M,dp ∆f MN2 i2 M,dn ∆f L (11) Cv Rv v2 Cv ∆f v2 Cv ∆f Rv Cv where • Γrms ≈ 1/2 for differential noise sources. • qmax is the total charge swing of the tank and is given by the monomial expression, qmax = Ctank Vamp = • Vsw . 2 Ltank ωres i2 M,dn ∆f MN1 i2 M,gn ∆f Ibias (12) P i2n /∆f is the sum of the current noise densities of the differential equivalent of individual noise sources: the transistor channel thermal noise (i2M,dT /∆f ), the transistor gate noise (i2M,gT /∆f ), the inductor thermal noise (i2RL /∆f ) and the varactor thermal noise (i2CV /∆f ), as depicted in Figure 4. i2 M,gn ∆f Figure 4: Noise sources in the LC oscillator. Inductor noise. There are two noise sources from the inductors: the ohmic losses in the winding and the losses in the substrate, i2RL 4kT ≈ 8kT =2· ∆f RL,p 1 1 + . Rp (Lω)2 /Rs (17) Since Rp , L and Rs are given by monomial expressions (see [12]), equation (17) is a posynomial function of the design variables. Varactor noise. The varactor noise can be modeled with the monomial expression, i2CV 4kT 8kT Cv ω =2· ≈ , ∆f Rv,p Qv (18) where Rv,p is the equivalent varactor parallel resistance. Since Γrms is a constant, qmax is monomial in the design variables and the noise sources i2n /∆f are posynomial in the design variables, expression (11) is a posynomial equation of the design variables. Therefore, we can minimize the phase noise at a given frequency. To guarantee a minimum phase noise over the entire tuning range, we can impose the constraint (L(foff ) ≤ Lmax (foff )) over several frequencies within the tuning range. Resonance frequency We can impose a constraint on the maximum resonance frequency ωres,max . This constraint together with a constraint on the tuning range is equivalent to specifying a center resonance frequency. The maximum tank resonance frequency is given by the posynomial expression: ωres,max = 1 . Ltank Ctank,min (19) We can therefore impose a minimum required ωres,max with the posynomial constraint ωres,max ≥ ωres,max,req . This constraint is always active (i.e., it is practically an equality). If it were not active the inductor could contribute additional capacitance to the tank, which would translate into a higher QL . Tuning range The tuning range is specified with two constraints Ltank Ctank,min Ltank Ctank,max 2 ≤ 1/ωmax ≥ 2 1/ωmin . (20) (21) Constraint (21) is not posynomial and cannot be handled directly by GP. Since constraint (20) is always tight at the optimum, we can handle constraint (21) indirectly. We rewrite constraint (21) as, 2 2 Ctank,max ≥ ωmax Ctank,min . ωmin 2 2 Now we let r = ωres,max /ωres,min , and obtain (r − 1) Ctank,min Cv,min +r ≤ 1. Cv,max Cv,max (22) Thus, we can substitute (21) by the posynomial constraint (22). Inductor constraints It is shown in [12] that many inductor specifications have the form of posynomial inequalities. For example, specifications on minimum and maximum inductance, on maximum area, on minimum turn spacing and width are monomial while specifications on minimum quality factor and minimum self-resonance frequency are posynomial. Geometry constraints We can require that the width and length of the devices be constrained within some range using monomial constraints, Wmin ≤ W ≤ Wmax Lmin ≤ L ≤ Lmax . (23) We can also limit the inductor area by imposing the monomial constraint d2out ≤ Amax . Loop gain For a complementary LC oscillator, the loop gain condition is (gm,n + gm,p ) /2 ≥ αg gtank , where αg is the excess gain and it is typically ≈ 2 − 3. Since we have gm,n = gm,p the loop gain condition is given by a posynomial constraint, gm,n ≥ αg gtank . Varactor tuning range Although the absolute value of the varactor capacitance is typically not limited, the maximum tuning ratio βv = Cv,max /Cv,min is limited. We impose a maximum βv with the monomial constraint Cv,max ≤ βv Cv,min . 6 Design examples In this section, we show two oscillator designs. The LC oscillators are designed in a standard 0.35µm five metal layer 2.5V CMOS process. Constraint Phase Noise 600kHz ωres Cload Itail Vsw Tuning range dout Excess loop gain αg Specification minimize 1.8GHz ≥ 0.2nF ≤ 5mA ≥ 2V ≥ 10% ≤ 300µm ≥3 Achieved −124.2dBc/Hz 1.8GHz 0.83nF 5mA 2.5V 10% 300µm 3 Table 1: Specifications for example 1. The objective in the first example is to minimize phase noise in an LC oscillator built using spiral inductors with PGS. The desired and achieved specifications are shown in Table 1. We note that many of the constraints are tight, i.e., power, voltage swing, tuning range and area are all set to the limit specified. Table 2 shows the optimal oscillator design obtained. It is important to note that significant phase noise is contributed by both the transistors (≈ 52%) and the inductors (≈ 44%). This shows that a design solely based on maximizing the inductor or its quality factor will not be optimum. In the second design example we limit Ibias to only 1.6mA. Table 3 shows the desired and achieved specifications. The differential swing voltage constraint is now tight. In table 4 we show the optimal design for the second example. The phase noise is around 7dB worse than in the first example and the relative contributions from the inductor and the transistors are ≈ 50% and ≈ 46%, respectively. This result is consistent with our intuition: increasing Ibias results in significant improvements in phase noise only if the oscillator operates in current limited mode. We cannot achieve a significant increase in the charge swing, qmax , in the voltage limited mode. Therefore, if the available power is low, it is necessary to to operate in deep current limited regime. This corresponds to smaller average transistor currents which results in smaller noise current density. Optimization of the inductor Q is more helpful in this case. Inductor 3.9nH 4 300µm 23.2µm 1.8µm 6.5Ω 0.3pF 332Ω Varactor Cv,max 0.81pF Cv,min 0.41pF Rv,p 4.2kΩ Transistors 17.0/0.35µm 52.7/0.35µm 0.07pF 0.21pF 2mS 5.5kΩ 1.7kΩ Tank Rtank 500Ω Ctank 1.1pF − 0.9pF qmax 2.3pC i2M,d /∆f 117.7pA2 /Hz L n dout w s Rs CL RL,p Wn /Ln Wp /Lp CNMOS CPMOS gm 1/gds,n 1/gds,p i2RL /∆f i2CV /∆f 99.2pA2 /Hz 7.9pA2 /Hz Table 2: Oscillator design for example 1. Constraint Phase Noise 600kHz ωres Cload Itail Vsw Tuning range dout Excess loop gain αg Specification minimize 1.8GHz ≥ 0.2nF ≤ 1.6mA ≥ 2V ≥ 10% ≤ 300µm ≥3 Achieved −118.5dBc/Hz 1.8GHz 0.2nF 1.6mA 2V 10% 300µm 3 Inductor 10.0nH 6 300µm 13.0µm 1.8µm 17.4Ω 0.24pF 804Ω Varactor Cv,max 0.32pF Cv,min 0.16pF Rv,p 10.7kΩ Transistors 6.8/0.35µm 21.0/0.35µm 0.03pF 0.08pF 2.4mS 15.8kΩ 5.0kΩ Tank Rtank 1250Ω Ctank 0.43pF − 0.36pF qmax 0.71pC i2M,d /∆f 38.0pA2 /Hz Wn /Ln Wp /Lp CNMOS CPMOS gm 1/gds,n 1/gds,p i2RL /∆f i2CV /∆f 41.0pA2 /Hz 3.1pA2 /Hz Table 4: Oscillator design for example 2. −121 Phase noise (dBc/Hz) In this paper, we have shown how a commonly used LC oscillator can be optimized rapidly and globally by posing the design problem as a geometric program. This formulation permits the designer to quickly explore globally optimal tradeoff curves between different specifications. The method allows the simultaneous optimization of all passive and active components. References [1] J. Craninckx and M. S. .J. Steyaert. A 1.8GHz CMOS low-phasenoise voltage-controlled oscillators with prescaler. IEEE Journal of Solid-State circuits, 30(12):1474–1482, December 1995. [2] A. Rofougaran et al. A 900MHz CMOS LC-oscillator with quadrature outputs. In IEEE International Solid-Sate Circuits conference, pages 392–393, 1996. [4] J. Craninckx and M. S. J. Steyaert. A 1.8GHz low-phase-noise CMOS VCO using optimized hollow spiral inductors. IEEE Journal of SolidState Circuits, 32(5):736–744, May 1997. [5] A. Hajimiri and T. H. Lee. Design issues in CMOS differential LC oscillators. IEEE Journal of Solid-State Circuits, 34(5), May 1999. [6] C. P. Yue, C. Ryu, J. Lau, T. H. Lee, and S. S. Wong. A physical model for planar spiral inductors on silicon. In Proceedings IEEE IEDM’96, 1996. [7] R. J. Duffin, E. L. Peterson, and C. Zener. Geometric Programming — Theory and Applications. Wiley, 1967. [8] Y. Nesterov and A. Nemirovsky. Interior-point polynomial methods in convex programming, volume 13 of Studies in Applied Mathematics. SIAM, Philadelphia, PA, 1994. [9] K. O. Kortanek, X. Xu, and Y. Ye. An infeasible interior-point algorithm for solving primal and dual geometric programs. Math Programming, 76:155–181, 1996. [11] C. P. Yue and S. S. Wong. On-chip spiral inductors with patterned ground shields for Si-based RF IC’s. IEEE Journal of solid-state circuits, 33(5):743–752, May 1998. −120 −122 [12] M. Hershenson, S. S. Mohan, S. P. Boyd, and T. H. Lee. Optimization of inductor circuits via geometric programming. In 36th Design Automation Conference, pages 994–998, 1999. −123 −124 −125 −126 20% 15% 10% −127 −128 2 4 6 8 10 12 Current (mA) Figure 5: Minimum phase noise versus power limit for different tuning ranges. 7 Conclusions [10] S. P. Boyd and L. Vandenberghe. Course Reader for EE364: Introduction to Convex Optimization with Engineering Applications. Stanford University, 1999. http://www.stanford.edu/class/ee364/. −119 −129 0 8 [3] B. Razavi. A 1.8GHz CMOS voltage-controlled oscillator. In IEEE International Solid-Sate Circuits conference, pages 388–389, 1997. Table 3: Specifications for example 2. L n dout w s Rs CL RL,p mal trade-off curves for the LC oscillator. In Figure 5 we show the trade-off curves of the minimum achievable phase noise versus quiescent power for different three tuning ranges. The rest of the specs are set to the values in table 2. As an RF designer might suspect, the phase noise decreases with higher power and smaller tuning range. What is difficult to know, without using the tool described in this paper, is exactly how these parameters trade off. Trade-off curves By repeatedly solving optimal design problems as we sweep over values of some of the specifications, we can obtain globally opti- [13] S. S. Mohan, M. Hershenson, S. P. Boyd, and T. H. Lee. Simple accurate expressions for planar spiral inductances. IEEE Journal of Solid-State Circuits, 34, October 1999. [14] A. Hajimiri. Current state of integrated oscillator design. In IEEE CSCC, 1999. [15] M. Hershenson, S. Boyd, and T. H. Lee. GPCAD: A tool for CMOS op-amp synthesis. In IEEE/ACM International Conference on Computer Aided Design, pages 296–303, San Jose, CA, 1998. [16] T. H. Lee. The design of CMOS radio-frequency integrated circuits. Cambridge University Press, 1998.